AI’s long solo tasks keep growing, but they are not yet a separate axis or a business — resolves 2027-07-23

How long a task a model can finish on its own keeps growing, measured by the research group METR as a time horizon, the length of task in human working time that a model completes at a given success rate, and the debate is whether this has become a separate axis of capability and something that sells.

If it has, vendors would price by task length or outcome, and companies would have to change how they buy and what their contracts say.

One side says the long-run axis has formed and is being commercialized, while SecondSource says the capability is unlocking but tracks four strands separately before drawing either conclusion.

SecondSource’s reading: models’ ability to complete long tasks on their own keeps unlocking, but that is not yet enough to say it has become a separate axis of capability or that it has been adopted commercially.

We track four strands separately, measurement, revenue, reliability and commercial signals, each settled on its own rather than rolled into one score.

The reliability baseline is METR’s May 2026 reading: at an 80% success rate, models finish tasks of about 3 to 4 hours, and we set ourselves 12 months to see whether that passes 8 hours.

In fairness: this card’s original sentence was lost from our records, and the current one was rebuilt from the four strands and reversal conditions already on file, adding no new call and changing neither the settlement method nor the deadline.

arXiv (Kwa/West/Becker...Barnes/Chan,METR)
“frontier AI time horizon has been doubling…”

BG2 Pod(Gerstner)× Atreides(Gavin Baker/Andrew Fox)× Altimeter(Clark Tang)
“there was a Copart tweet about this yesterday. He…”

Dwarkesh Podcast
“if you look at the size of a Tesla, and if you…”

Dwarkesh Podcast (Sholto Douglas, then at Google Gemini and later at Anthropic; Trenton Bricken, Anthropic interpretability)
“I think that's more about nines of reliability and…”

Latent Space / AINews(swyx)
“GPT-5.6 Sol had a detected cheating rate higher…”

Latent Space / AINews(swyx)
“RL on health-domain conversations reinforcing…”

Settled in four legs: METR's successor to TH1.1 on the 50% horizon (stalling two or more doubling periods weakens the axis claim); Anthropic or OpenAI revenue breakdowns isolating long-horizon agent products; whether the 80% horizon passes 8 hours within 12 months (baseline 3-4 hours, May 2026); any frontier lab writing duration into pricing or SLAs.

What would overturn it

A lab shipping duration- or outcome-priced model SKUs strengthens the axis-shift reading; the 80% horizon not passing 8 hours in 12 months downgrades the unlock; revenue breakdowns showing long-horizon products above half unblocks the causal claim; a stalled 50% horizon weakens the axis and favors the view that capability is already sufficient.